Self-supervised cloud semantic segmentation with vision transformers

vision transformers trained without explicit supervision based on the DINO framework from https://arxiv.org/abs/2104.14294

applied to MODIS satelite images of derived cloud properties:

• https://ladsweb.modaps.eosdis.nasa.gov/missions-and-measurements/products/MOD06_L2
• https://ladsweb.modaps.eosdis.nasa.gov/missions-and-measurements/products/MYD06_L2

and to level 1b radiances:

...

workflow

1. Download the raw MODIS from NASA with login

2. Reproject to uniform lat-lon grid (for future climate model compatibility)

3. Engineer training stacks, normalize, etc a. liquid water path, ice water path, cloud top pressure b. RGB c. some other bands?

4. Fit vanilla ViT

open questions:

• how many heads in last layer (implicit number of classes)?

• how to scale to 2kx1.3k pixel images?

• how do do sub-patch (in the ViT sense) level classification

....